May 11ByteDancetechnology team announced that the new LangStack-based Deep Research Project - DeerFlow OfficialOpen Source.
According to the official introduction, DeerFlow supports features including in-depth research, MCP integration, report AI-enhanced editing, and podcast generation. It is worth mentioning that DeerFlow supports the popular Replay mode (i.e., the process of multiple rounds of interaction with a large model is restored in the form of a quick replay).
DeerFlow is designed on a new Multi-Agent architecture with a unique Research Team mechanism that supports multiple rounds of dialog, decision-making and task execution.
At the same time, DeerFlow supports the generation of two-host podcasts from reports, with the voice technology of the Volcano engine, and rich timbre, you can generate very natural podcast audio content. In addition, DeerFlow also supports generating PPT from reports, and supports generating text version of PPT.
DeerFlow is currently available on the official website (https://deerflow.tech/) and the ByteHop GitHub repository (https://github.com/bytedance/deer-flow). It is officially recommended to run DeerFlow with Beanbag 1.5 Pro.
In addition, ByteDance also recently open-sourced the Seed 8.2B parameter scale code model "Seed-Coder".
Seed-Coder has three versions, "Base", "Instruct" and "Reasoning", and is based on "doubao-coder" with Llama 3 structure.
It is worth noting that Seed-Coder focuses on "minimal human involvement" and improves the model code programming ability by generating and filtering high-quality training data itself.
The Instruct version performed well in the Programming test, taking two benchmark SOTA's, while the Reasoning version outperformed QwQ-32B and DeepSeek-R1 at IOI 2024.
Seed-Coder is also now available on the official website (https://bytedance-seed-coder.github.io/).